on the energy efficiency prospects of network … the energy efficiency prospects of network...
TRANSCRIPT
On the Energy Efficiency Prospects of NetworkFunction Virtualization
Rashid Mijumbi†, Joan Serrat†, Juan-Luis Gorricho† and Javier Rubio-Loyola‡†Universitat Politecnica de Catalunya, 08034 Barcelona, Spain
‡CINVESTAV, Tamaulipas, Mexico
Abstract—Network Function Virtualization (NFV) has re-cently received significant attention as an innovative way ofdeploying network services. By decoupling network functionsfrom the physical equipment on which they run, NFV hasbeen proposed as passage towards service agility, better time-to-market, and reduced Capital Expenses (CAPEX) and OperatingExpenses (OPEX). One of the main selling points of NFV is itspromise for better energy efficiency resulting from consolidationof resources as well as their more dynamic utilization. However,there are currently no studies or implementations which attachvalues to energy savings that can be expected, which could makeit hard for Telecommunication Service Providers (TSPs) to makeinvestment decisions. In this paper, we utilize Bell Labs’ GWATTtool to estimate the energy savings that could result from thethree main NFV use cases−Virtualized Evolved Packet Core(VEPC), Virtualized Customer Premises Equipment (VCPE) andVirtualized Radio Access Network (VRAN). We determine thatthe part of the mobile network with the highest energy utilizationprospects is the Evolved Packet Core (EPC) where virtualizationof functions leads to a 22% reduction in energy consumption anda 32% enhancement in energy efficiency.
Keywords—Network function virtualization, cloud computing,energy efficiency, estimation, measurements.
I. INTRODUCTION
TSPs have been facing an explosion in the data that theirnetworks must support. And this will not stop soon. Accordingto Cisco [1], the annual global IP traffic will nearly triple from2014 to 2019. Specifically, the traffic will pass the zettabyte1
threshold by the end of 2016, and reach 2 zettabytes peryear by 2019, at which time two-thirds of all this traffic willoriginate from non-PC devices. With this traffic burst, TSPsmust correspondingly expand and maintain their networksboth of which lead to increased CAPEX and OPEX. Yet,due to competition both among themselves and from servicesprovided over-the-top on their infrastructure, TSPs cannotrespond to these increased costs by increasing subscriptionfees. This has lead to reductions in TSP revenues [2], andforced them to search for ways of deploying and maintainingnetworks at reduced costs.
NFV [3], [4] has been proposed as a viable path towardsachieving this goal. By leveraging recent advances in virtu-alization technology, the main idea of NFV is to decoupleNetwork Functions (NFs) from the physical equipment onwhich they run. This way, the NFs can be virtualized, andrun on industry standard servers and switches which couldbe located in either data centers or network nodes. The
1A zettabyte is equal to 1000 exabytes, which is equal to 1000 billiongigabytes.
Virtual Network Functions (VNFs) may then be relocated andinstantiated in different network locations (for example aimedat introduction of a service targeting customers in a givengeographical location) without necessarily requiring purchaseand installation of new hardware [5].
One of the network parts which is expected to benefitfrom NFV is the EPC, which is the core network for LongTerm Evolution (LTE) as specified by 3GPP [6]. In Fig.1, we show a basic architecture of LTE without NFV. TheUser Equipment (UE) is connected to the EPC over the LTEaccess network (E-UTRAN). The evolved NodeB (eNodeB) isthe base station for LTE radio. The EPC performs essentialfunctions including subscriber tracking, mobility managementand session management. It is made up of four NFs: ServingGateway (S-GW), Packet Data Network (PDN) Gateway (P-GW), Mobility Management Entity (MME), and Policy andCharging Rules Function (PCRF). It is also connected toexternal networks, which may include the IP Multimedia CoreNetwork Subsystem (IMS). In current EPC implementations,all its functions are based on proprietary equipment spreadacross the core network. Therefore, even minor changes to agiven function may require a replacement of the equipment.The same applies to cases when the capacity of the equipmenthas to be changed.
Fig. 2, shows the same architecture in which the EPC isvirtualized. In this case, either all functions in the EPC, or onlya few of them are transferred to a shared (cloud) infrastructure.Virtualizing the EPC could potentially lead to better flexibilityand dynamic scaling, and hence allow TSPs to respond easilyand cheaply to changes in market conditions. For example, asrepresented by the number of servers allocated to each functionin Fig. 2, there might be a need to increase user plane resourceswithout affecting the control plane. In this case, VNFs suchas a virtual MME may scale independently according totheir specific resource requirements. In the same way, VNFsdealing with the data plane might require a different numberof resources than those dealing with signaling only. Thisflexibility is expected to lead to more efficient utilization ofresources. Finally, it also allows for easier software upgradeson the EPC network functions, which would hence allow forfaster launch of innovative services.
Breaking the bond between functions and physical equip-ment is expected to lead to several advantages. First, Sincethe NF can be deployed, chained and updated remotely, thispromises more flexibility, agility and reduced time-to-marketfor services. In addition, as it would no longer be necessary tocompletely change network equipment as network technologieschange (to accommodate the data requirements), CAPEX could
arX
iv:1
512.
0021
5v1
[cs
.NI]
1 D
ec 2
015
Access Network (E‐UTRAN)
User Equipment
Evolved Packet Core (EPC)
eNodeB eNodeB S‐GW MME
PCRF P‐GW
External Networks
Fig. 1. Traditional LTE EPC
Access Network (E‐UTRAN)
User Equipment
VNFs
eNodeB eNodeBS‐GW
MME
PCRF
P‐GW
External Networks
Data Centers
Fig. 2. Possible Virtualized EPC Implementation
be significantly reduced. Finally, as a result of deploying NFon virtualized resources, NFV promises the ability to scaleresource allocations up and down as traffic demands ebb andflow. This is expected to potentially reduce the number ofphysical devices operating at any point, and hence reduce TSPsenergy bills. Since energy bills represent more than 10% ofTSPs’ OPEX2 [7], reduced energy consumption is one of thestrong selling points of NFV. This means that energy efficiencywill be critical key performance indicator for NFV.
In this paper, we use Bell Lab’s GWATT tool [7] to estimatethe energy savings that could result from the three main usescases (VEPC, VRAN and VCPE [8]) of NFV. The tool usesstate-of-the-art network and traffic models to estimate not onlythe expected changes in user traffic, but also the resulting effecton network equipment in terms of energy consumption. It alsoincludes possibilities to set different network architectures, in-cluding possibilities to select virtualized networks. We believethat an understanding of the actual savings in energy that canbe derived from virtualizing different parts of the network isvery important in guiding TSPs to make investment decisions.In particular, at this early stage of NFV where virtualizedfunctions will exist along side those running on specializedhardware, it could be important for TSPs to determine whichpart of the network could be virtualized first based on theexpected gains.
The rest of this paper is organized as follows: sectionII discusses related work, while the estimation tool used indescribed in detail in section III. In section IV, we presentand discuss results from utilizing the tool, before concludingthe paper in section V.
II. RELATED WORK
China Mobile recently published [9] their experiences indeploying a Cloud Radio Access Network (C-RAN). Oneof the tests was performed on their 2G and 3G networks,where it was observed that by centralizing the RAN, powerconsumption could be reduced by 41% due to shared air-conditioning. In addition, Shehab et al. [10] analyzed thetechnical potential for energy savings associated with shiftingU.S. business software to the cloud. The results suggested asubstantial potential for energy savings. In fact, the authorsnoted that if all U.S. business users shifted their email,productivity software, and CRM software to the cloud, the
2This figure can raise up to 45% for TSPs in less developed countries[7].
primary energy footprint of these software applications couldbe reduced by as much as 87%.
DROP [11] is a middleware platform which was originallyaimed at creating software routers on top of commodityservers, while hiding the complexity of its modular architectureto control-plane applications and system administrators. It hasbeen recently extended to DROPv2 [12] which is focusedon implementing more efficient power management in SDN.The idea of DROPv2 is to periodically calculate, based ontotal traffic, the number of forwarding elements that should beput into a standby state, and how the remaining ones shouldshare the available traffic. Both these research activities do notspecifically look at virtualized NFs.
Yathirah et al. [13] identified a need to extend Openstack toinclude a scheduler which is energy-aware. The idea is to useanalytics to determine the current status, and use it to scheduleOpenstack resources. However, the authors do not give anydetails about solution or the actual resulting energy savings.Instead of relying on generic servers, [14] uses ApplicationSpecific Instruction-set Processor (ASIP) to achieve an energy-efficient deployment of a virtualized deep packet inspection(DPI). Based on experiments, the authors argue that theyare able to achieve superior energy efficiency compared toDPIs deployed on commodity servers. However, this wouldbe against the general concept of NFV as it runs the functionson specialized servers. In addition, it only considers a specific,light function. Results could be different on a large scale, formore processing intensive functions or chains of functions.
To the best of our knowledge, the work presented in thispaper is the first study and discussion of the forecast energyefficiency for any of the NFV use cases. Given the importanceof energy efficiency as a key performance indicator for NFV,we hope that the results of this study can help TSPs makenecessary investment decisions with regard to the evolutiontowards NFV. We also hope that this can direct researcherstowards proposing energy-aware algorithms to those parts ofthe network where savings may be highest. It is importantto note that the GWATT tool used in this paper has NOTbeen developed by the authors. Our contribution is in usingthe tool−which is freely available, and aimed at determiningenergy efficiency in multiple network scenarios−to study,simulate, summarize and discuss results specific to chosenNFV use cases.
Fig. 3. Network Details from GWATT for Virtualized EPC
TABLE I. SUMMARY OF RESULTS FOR THE BASELINE NETWORK
DomainHome & Enterprise 59.85 0.012807 14,422.00 Access & Aggregation 62.36 0.009887 27,501.50 Metro 163.07 0.760441 661.85 Edge 83.26 0.325516 789.48 Core 166.53 1.770110 290.37 Service Core & Data Centre 381.47 0.023521 50,057.60 Home & Enterprise 46.87 0.012615 11,466.40 2,955.60 Access & Aggregation 62.36 0.009887 27,501.50 - Metro 221.14 0.625239 1,091.62 (429.77) Edge 83.26 0.325516 789.48 - Core 166.53 1.770110 290.37 - Service Core & Data Centre 381.47 0.023521 50,057.60 - Home & Enterprise 59.85 0.012807 14,422.00 - Access & Aggregation 62.04 0.066784 10,900.00 16,601.50 Metro 221.72 0.591841 1,156.23 (494.38) Edge 83.26 0.325516 789.48 - Core 166.53 1.770110 290.37 - Service Core & Data Centre 381.47 0.023521 50,057.60 - Home & Enterprise 59.85 0.012807 14,422.00 - Access & Aggregation 62.36 0.009887 27,501.50 - Metro 163.07 0.760441 661.85 - Edge 79.80 0.325516 756.66 32.82 Core 166.53 1.770110 290.37 - Service Core & Data Centre 384.93 0.040126 29,608.70 20,448.90
Baseline 916.54 0.033012 93,722.80 -
Traffic(Exabytes per month) Total Efficiency (MBITS/J) Total Power (MWATTS) Power Savings (MWATTS)
VRAN 974.86 0.043241 77,615.68 16,107.12
VCPE 961.63 0.035689 91,196.97 2,525.83
VEPC 916.54 0.043381 73,241.07 20,481.72
III. BELL LABS’ GWATT
Bell Labs’ GWATT tool [7] is an important step in de-termining the possible effect, on energy consumption, of theevolution to NFV. Based on some forecast for traffic growth,the tool is able to show the effect of virtualizing differentnetwork functions. It also gives a cummulative forecast onenergy savings over a five year period. As shown in Fig. 3, thetool divides the network into six domains (Home & Enterprise,Access & Aggregation, Metro, Edge, Core and Service Coreand Data Centers) which are defined below.
The home & enterprise is made up of consumer CustomerPremises Equipment (CPE) in the home, such as DSL routers,set-top boxes and cable modems. It also includes the network-ing equipment (LAN switches, routers) used within enterprisesites. The access and aggregation is used to connect users tothe network using fixed or wireless access, aggregates wirelesssites, and provides public switched telephone network (PSTN)service. It includes macro cells each with a tower, antenna,radio and network equipment. It connects multiple wireless cellsites to the mobile operator’s network, typically using point-to-point connections. The metro provides aggregation, transportand traffic engineering for residential, business and wirelessservices. It connects subscribers to a larger service networkor the Internet. The edge serves as the primary entry point
to service provider core networks and runs service routing,intelligence and signaling functions. The core provides highperformance IP routing and optical transport for the serviceprovider’s backbone network and internet connection. It in-cludes optical transport systems which can support multiple10 Gbps channels on each fiber and provide the high speedoptical backbone to transport IP core traffic. Finally, the servicecore and data center domain consists of telecommunicationoperational systems and the data center equipment to host cus-tomer content and run customer applications. The service coreconsists of multiple telecom service delivery platforms andoperational IT systems required to deliver telecommunicationsservices and run the business.
Each of the six network domains can be edited to selectdifferent network models and technologies and hence analyzeits energy impact. As shown in the top left corner of Fig. 3,the figure is from an analysis of the energy consumption ofa network in which the EPC is virtualized. The results fromthe analysis are shown at the bottom of the figure. To discussthese results, we start by noting that GWATT is based on twomain models:
1) A traffic data model. For each domain of thenetwork, the tool uses a traffic volume expressedin exabytes/month. The traffic volume is based on
TABLE II. SUMMARY OF RESULTS FOR VEPC, VRAN AND VCPE
DomainHome & Enterprise 46.87 0.012615 11,466.40 2,955.60 Access & Aggregation 62.36 0.009887 27,501.50 - Metro 221.14 0.625239 1,091.62 (429.77) Edge 83.26 0.325516 789.48 - Core 166.53 1.770110 290.37 - Service Core & Data Centre 381.47 0.023521 50,057.60 - Home & Enterprise 59.85 0.012807 14,422.00 - Access & Aggregation 62.04 0.066784 10,900.00 16,601.50 Metro 221.72 0.591841 1,156.23 (494.38) Edge 83.26 0.325516 789.48 - Core 166.53 1.770110 290.37 - Service Core & Data Centre 381.47 0.023521 50,057.60 - Home & Enterprise 59.85 0.012807 14,422.00 - Access & Aggregation 62.36 0.009887 27,501.50 - Metro 163.07 0.760441 661.85 - Edge 79.80 0.325516 756.66 32.82 Core 166.53 1.770110 290.37 - Service Core & Data Centre 384.93 0.040126 29,608.70 20,448.90
Traffic(Exabytes per month) Total Efficiency (MBITS/J) Total Power (MWATTS) Power Savings (MWATTS)
VRAN 974.86 0.043241 77,615.68 16,107.12
VCPE 961.63 0.035689 91,196.97 2,525.83
VEPC 916.54 0.043381 73,241.07 20,481.72
Home & Enterprise
0
10,000
20,000
30,000
40,000
50,000
60,000
70,000
80,000
90,000
100,000
Baseline VCPE VRAN VEPC
Po
wer
Co
nsu
mp
tio
n (
MW
ATT
S)
Fig. 4. Total Power
0
5000
10000
15000
20000
25000
VCPE VRAN VEPC
Po
wer
Sav
ings
(M
WA
TTS)
Fig. 5. Power Savings
0
5
10
15
20
25
30
35
40
-10% 0% 10% 20% 30% 40% 50%
Po
wer
Sav
ings
(M
WA
TTS)
Tho
usa
nd
s
Traffic Growth Rate
Virtual EPC
Virtual CPE
Virtual RAN
Fig. 6. Variation of Power Savings
extrapolated traffic projections performed by BellLabs and based on real data. For example, as shownin Figure 3, when the EPC is virtualized, the trafficassociated to the network core domain is 166.53exabytes/month, while that associated with the datacenter and service core is 384.93 exabytes/month.
2) A network element efficiency model. An efficiencyexpressed in megabits/joules is associated with eachdomain. For instance, after virtualizing the EPC asshown in Figure 3, it can be observed that thepower efficiency for the core network domain is1.77011 megabits/joules while that for the data centerand service core domain is 0.04013 megabits/joules.We also note that with this efficiency, the totalpower consumed by all network elements in each ofthese domains is 290.36 Megawatts and 29, 608.69Megawatts respectively.
Based on the traffic and efficiency data above, and takinginto consideration the underlying network model, GWATTcomputes the total end-to-end network power (expressed inmega watts), energy consumption (in gigajoules). In addition,by comparing all these values with those from a baselinenetwork3, the tool is able to determine the forecast savings inenergy by changing a given part of the end-to-end network. Itcan be observed that by virtualizing the EPC, 32.82 Megawattsof power are saved at the edge while 20, 442.95 Megawatts
3A baseline network is one where all functions are run in physicalequipment, using the tool’s default technologies and settings.
are saved in the data center and service core. These traffic,consumption and efficiency values for the baseline networkare shown in Table I.
In order to use the tool, some choices need to be made4.These settings relate to the network size, traffic growth rate,number of applications and network architecture. The defaultnetwork size worldwide, but this can be set to one of NorthAmerica, Small Country, Large Country etc. The tool alsoallows one to set the rate at which traffic grows in theselected network size as a percentage. The traffic applicationssetting may be used to select one or more applications ofthe available application which include online gaming, filesharing, video, web, business video, business file sharing andbusiness web. The default setting utilizes all the applications,which represents 100% of the set traffic. Finally, it is possibleto select a particular network architecture. While the toolincludes many architectures, the focus of this paper is onthe virtualization of the CPE, RAN and EPC. The VEPCvirtualizes mobile network functions by implementing a cloud-based EPC architecture that uses resources more efficiently andimproves agility and scale. The VRAN virtualizes and cen-tralizes the consolidation of eNodeB Baseband Units (BBU)from multiple cell sites to improve scale and efficiency. Finally,the VCPE virtualizes the residential gateway of the CPE toreduce complexity, energy consumption and deployment cost(CAPEX).
4Unless stated otherwise, the results presented in this paper are based onGWATT’s default settings.
0
10,000
20,000
30,000
40,000
50,000
60,000
Baseline VCPE VRAN VEPC
Po
wer
Co
nsu
mp
tio
n (
MW
ATT
S)
Fig. 7. Power Consumption of Service Core
0
5,000
10,000
15,000
20,000
25,000
30,000
Baseline VCPE VRAN VEPC
Po
wer
Co
nsu
mp
tio
n (
MW
ATT
S)
Fig. 8. Power Consumption of Access Network
0
2,000
4,000
6,000
8,000
10,000
12,000
14,000
16,000
Baseline VCPE VRAN VEPC
Po
wer
Co
nsu
mp
tio
n (
MW
ATT
S)
Fig. 9. Power Consumption of Home Network
880
890
900
910
920
930
940
950
960
970
980
Baseline VCPE VRAN VEPC
Tota
l Tra
ffic
(Ex
abyt
es p
er m
on
th)
Fig. 10. Total Traffic
0
50
100
150
200
250
300
350
400
450
Bas
elin
e
VC
PE
VR
AN
VEP
C
Bas
elin
e
VC
PE
VR
AN
VEP
C
Bas
elin
e
VC
PE
VR
AN
VEP
C
Bas
elin
e
VC
PE
VR
AN
VEP
C
Bas
elin
e
VC
PE
VR
AN
VEP
C
Bas
elin
e
VC
PE
VR
AN
VEP
C
Home &Enterprise
Access &Aggregation
Metro Edge Core Service Core &Data Centre
Traf
fic
Dis
trib
uti
on
(Ex
abyt
es p
er m
on
th)
Fig. 11. Traffic Distribution
0.0000
0.0050
0.0100
0.0150
0.0200
0.0250
0.0300
0.0350
0.0400
0.0450
0.0500
Baseline Virtual EPC Virtual CPE Virtual RAN
Efficiency (MBITS/J)
Fig. 12. Efficiency
IV. RESULTS
The results for the three NFV use cases are summarized inTable II. The power savings are determined in comparison withthe results from the baseline network in I. These results, in-cluding others with different setting are presented in graphicalform in Figs. 4 - 12.
From Fig. 4 and 5, it can be observed virtualization ofnetwork functions leads to a reduced power consumption.However, it can be noted that VEPC achieves the highestreduction in power consumption, followed by VRAN and thenVCPE. The reason for superiority of virtualizing the EPC couldbe due to the higher traffic that goes through the EPC domainas shown in Table I. This means that there are many networkelements in the core, which presents more possibility to reduceenergy not only by consolidating these network elements,but by also reducing the required capacity through the betterresource management presented by virtualization. In fact, thiscan be further proved by looking at Fig. 6 which shows how thepower savings change if the rate of traffic growth is increased.It can be noted while the VEPC saves 21% of power evenwhen traffic is reducing, the savings increase to about 23% inconditions of traffic growing at 50%.
In Figs. 7, 8 and 9, we show how virtualizing a givennetwork domain affects the energy consumption of the otherdomains. As expected, it can be observed that each of thethree use cases considered have the strongest effect in termsof power savings on the respective domain which has beenvirtualized.
In Fig. 10, we show the total network traffic carried ineach use case, which is broken up across all the contributingnetwork domains in Fig. 11. As earlier noted, it is evident thatthe service core and data center domain has the highest trafficin all cases. However, it is interesting to note that while this
traffic is almost constant across all the other network domainsfor the 4 cases evaluated (including the baseline), it variesconsiderably in the metro domain. In fact, the total traffic trafficin Fig. 10 takes on a similar profile to the traffic in the metronetwork. This could mean that the virtualization of any of theconsidered use cases could benefit more from virtualizationof the metro network. The idea is that if the traffic increases,there is potential to benefit for improved resource managementthrough virtualization of resources.
Finally, 12 shows the total efficiency resulting from vir-tualizing of the considered domains. Just like the savings inpower consumption, and probably for the same reasons, theEPC benefits more from virtualization than both the RAN andCPE. The fact that the CPE lags in terms of energy efficiencyand/or power savings in all these results is not so surprising.CPEs normally consume little energy and the advantages fromvirtualizing the CPE could be much more on the side ofCAPEX (as the boxed would become cheaper on a large scale).The OPEX savings will mainly be due to ability to makeremote configurations as well as upgrades.
However, while the results presented in this paper showpromising values on energy savings, a lot of work will likelyneed to be done to achieve them in practice. The fact that NFVmakes data centers an indispensable part of telecommunicationnetworks raises questions on the actual value of savings inenergy consumption (if any) that will be achieved. Accordingto an analysis in the SMARTer 2020 report from GeSI [15],the total electricity requirements of the cloud (including datacenters and networks, but not devices) in 2011 was 684 billionkWh. The cloud, if it were a country, would rank 6th in theworld in terms of its energy demand, and yet this demandis expected to increase by 63% by 2020 [16]. While someprogress on energy efficient cloud computing has been made,the fast growing energy needs of data centers continue to
receive a lot of attention [17], [18]. Therefore, there is anurgent need to evaluate whether NFV will be able to meet theseenergy saving expectations, or whether−like the NFs−theenergy consumption will just be transferred to the cloud.
V. CONCLUSION
In this paper, we have attached values to the energysaving expectations of some NFV use cases. Using Bell Lab’sGWATT tool, we have determined that by virtualizing the EPC,RAN and CPE we can reduce network power consumption by22%, 17% and 3% respectively in comparison to a baselineworldwide network consumption of 93,722.80 Megawatts. Wehave also shown that this saving in energy consumption couldeven be higher in cases of increasing traffic across the network.Finally, we have noted that virtualizing each of these net-work domains does not significantly affect the traffic passingthrough the domain, instead affecting the traffic through themetro network. For this reason, its important to consider thevirtualization of the metro network as a necessary aspect ofvirtualizing other network domains.
However, while GWATT is an important step in attachingnumbers to the energy savings expected from NFV, it canstill be improved. In particular, it currently does not have adetailed technical documentation. In addition, Cisco’s visualnetworking index [1] forecasts that annual global IP trafficwill reach 1000 exabytes in 2016. Based on this, the (monthly)traffic values in Tables I and II seem to be so high, yet thetool does not provide a possibility to know how these valuesare derived. Without a doubt, the energy efficiency of cloudbased NFs will continue to receive attention. NFV will putInPs under even more pressure to manage energy consumptionnot to only to cut down energy expenses, but also to meetlegal, regulatory and environmental requirements. Therefore,approaches for placing network functions [19], [20] as wellas those that dynamically manage the network resources [21],[22], [23], [24] must be geared towards energy efficiency justas they are focused on efficient resource utilization.
ACKNOWLEDGMENT
This work is partly funded by FLAMINGO, a Networkof Excellence project (318488) supported by the EuropeanCommission under its Seventh Framework Programme, andproject TEC2012-38574-C02-02 from Ministerio de Economiay Competitividad.
REFERENCES
[1] Cisco, “The Zettabyte Era: Trends and Analysis,” Tech. Rep., May 2015.[2] R. Mijumbi, J. Serrat, J.-L. Gorricho, N. Bouten, F. De Turck, and
R. Boutaba, “Network function virtualization: State-of-the-art and re-search challenges,” 2015.
[3] R. Guerzoni, “Network Functions Virtualisation: An Introduction, Ben-efits, Enablers, Challenges and Call for Action. Introductory whitepaper,” in SDN and OpenFlow World Congress, June 2012.
[4] R. Mijumbi, J. Serrat, J.-L. Gorricho, S. Latre, M. Charalambides,and D. Lopez, “Management and orchestration challenges in networkfunction virtualization,” 2015.
[5] R. Mijumbi, J. Serrat, J.-L. Gorricho, N. Bouten, F. De Turck, andS. Davy, “Design and Evaluation of Algorithms for Mapping andScheduling of Virtual Network Functions,” in IEEE Conference onNetwork Softwarization (NetSoft). University College London, April2015.
[6] “The 3rd Generation Partnership Project (3GPP),” http://www.3gpp.org/about-3gpp/about-3gpp, 2015, Accessed: February, 10 2015.
[7] Bell Labs, Alcetel Lucent, “G.W.A.T.T. (Global What if Analyzer ofNeTwork Energy ConsumpTion). Bell Labs application able to measurethe impact of technologies like SDN & NFV on network energyconsumption. White Paper.” http://gwatt.net/intro/1, 2015.
[8] ETSI Industry Specification Group (ISG) NFV, “ETSI GSNFV 001 V1.1.1: Network Function Virtualization. Use Cases,”www.etsi.org/deliver/etsi gs/NFV/001 099/001/01.01.01 60/gsNFV001v010101p.pdf, October 2013.
[9] C.-L. I, J. Huang, R. Duan, C. Cui, J. Jiang, and L. Li, “Recent Progresson C-RAN Centralization and Cloudification,” Access, IEEE, vol. 2, pp.1030–1039, September 2014.
[10] E. Masanet, A. Shehabi, L. Ramakrishnan, J. Liang, X. Ma, B. Walker,V. Hendrix, and P. Mantha, “The Energy Efficiency Potential of Cloud-Based Software: A US Case Study. Lawrence Berkely National Labo-ratory, Berkerly California.” 2013.
[11] R. Bolla and R. Bruschi, “An open-source platform for distributed linuxsoftware routers,” Comput. Commun., vol. 36, no. 4, pp. 396–410, Feb.2013. [Online]. Available: http://dx.doi.org/10.1016/j.comcom.2012.11.002
[12] R. Bolla, C. Lombardo, R. Bruschi, and S. Mangialardi, “DROPv2:energy efficiency through network function virtualization,” Network,IEEE, vol. 28, no. 2, pp. 26–32, March 2014.
[13] Yathiraj, Udupi and Dutta Debo and Krishnan, Ram, “IncreasingNetwork and Energy Efficiency via Optimized NFV Placement inOpenstack Clouds,” OpenStackSummit,Atlanta, May 2014.
[14] K. Yamazaki, T. Osaka, S. Yasuda, S. Ohteru, andA. Miyazaki, “Accelerating sdn/nfv with transparent offloadingarchitecture,” in Presented as part of the Open NetworkingSummit 2014 (ONS 2014). Santa Clara, CA: USENIX,2014. [Online]. Available: https://www.usenix.org/conference/ons2014/technical-sessions/presentation/yamazaki
[15] “Global e-Sustainability Initiative (GeSI) SMARTer2020,” http://gesi.org/SMARTer2020, 2015, Accessed: June, 08 2015.
[16] GREENPEACE, “Clicking Clean: How Companies are Creating theGreen Internet,” www.greenpeace.org, April 2014.
[17] The Natural Resources Defense Council (NRDC), “Data Center Ef-ficiency Assessment. Scaling Up Energy Efficiency Across the DataCenter Industry: Evaluating Key Drivers and Barriers. Issue Paper.”August 2014.
[18] A. Beloglazov, R. Buyya, Y. C. Lee, and A. Y. Zomaya, “ATaxonomy and Survey of Energy-Efficient Data Centers and CloudComputing Systems.” Advances in Computers, vol. 82, pp. 47–111,2011. [Online]. Available: http://dblp.uni-trier.de/db/journals/ac/ac82.html#BeloglazovBLZ11
[19] R. Mijumbi, J. Serrat, J.-L. Gorricho, and R. Boutaba, “A path genera-tion approach to embedding of virtual networks,” Network and ServiceManagement, IEEE Transactions on, vol. 12, no. 3, pp. 334–348, 2015.
[20] R. Mijumbi, J.-L. Gorricho, and J. Serrat, “Contributions to efficientresource management in virtual networks,” in Monitoring and SecuringVirtualized Networks and Services. Springer Berlin Heidelberg, 2014,pp. 47–51.
[21] R. Mijumbi, J.-L. Gorricho, J. Serrat, M. Claeys, F. De Turck, andS. Latre, “Design and evaluation of learning algorithms for dynamicresource management in virtual networks,” in Network Operations andManagement Symposium (NOMS), 2014 IEEE. IEEE, 2014, pp. 1–9.
[22] R. Mijumbi, J.-L. Gorricho, J. Serrat, M. Shen, K. Xu, and K. Yang, “Aneuro-fuzzy approach to self-management of virtual network resources,”Expert Systems with Applications, vol. 42, no. 3, pp. 1376–1390, 2015.
[23] R. Mijumbi, J. Serrat, J. Rubio-Loyola, N. Bouten, F. De Turck, andS. Latre, “Dynamic resource management in sdn-based virtualizednetworks,” in Network and Service Management (CNSM), 2014 10thInternational Conference on. IEEE, 2014, pp. 412–417.
[24] R. Mijumbi, J.-L. Gorricho, J. Serrat, M. Claeys, J. Famaey, andF. De Turck, “Neural network-based autonomous allocation of resourcesin virtual networks,” in Networks and Communications (EuCNC), 2014European Conference on. IEEE, 2014, pp. 1–6.